In short: your words never reach our servers — there’s nothing in the architecture for them to reach. They go only where you point BreatheReply: the provider behind your own API key, or a model running on your own machine. Pick a local model and zero tokens leave your device. Everything below is what those three paths actually look like.
A reply is one of the most private things you write all day.
It isn’t a search query or a rough draft. It’s you talking to one specific person — your boss, your client, your partner, your mother — usually about something delicate, and almost never something you’d post where anyone could read it.
So before we talk about anything else, it’s fair to ask: when a tool helps you write one of those, where does that text go?
Where do your words go when AI helps you reply?
There are three common answers, and they are not remotely equivalent:
| Paste into a cloud chatbot | Bring your own key | Run a local model | |
|---|---|---|---|
| Where the text goes | That company’s servers | The provider you chose | Nowhere — your own machine |
| Who sees the text | That company’s operators | The provider you chose — not us | Nobody outside your device |
| Who sets the retention rules | Their policy, which you don’t negotiate | Your account, your provider’s terms | You |
| Does the network get used at all | Yes | Yes | No |
The middle column is what most AI writing tools quietly leave unclear. “Bring your own key” gets used as a feature bullet without explaining what actually changes. So here it is plainly: when the app talks directly to a provider using your key, there is no company sitting in the middle holding a copy of your conversations. The relationship is yours. We’re not party to it.
That distinction is the whole product, and it’s easy to blur in a sentence. Let’s not.
What changes when you bring your own key?
Three things, concretely:
- Your words never reach us. Not “we delete them after thirty days,” not “anonymized.” We don’t have a copy because there is no component of ours in the path your messages travel. There’s nothing here for us to store, which is why “we never store your messages” is a statement about architecture rather than a policy we promise to honour. Storage implies something arrived; nothing does. The full statement lives in our privacy policy.
- Your key stays on this device. It’s never baked into the installer and never synced anywhere. Once you save it, the app won’t display it again — so keep your own copy somewhere safe. Windows encrypts it using credentials tied to your own account, which means it can’t be opened by a different user on the same machine, and it won’t survive being copied elsewhere.
- You decide when the app is allowed to talk. An unconfigured app with no local model does nothing at all. There’s no relay of ours in the path your messages take — no fallback model we quietly call on your behalf when your own key isn’t set up. If the app can’t reach a model you chose, it says so.
And if even that feels like too much exposure, there’s the third column.
What about the “never leaves your machine” claim?
This is the phrase worth being precise about, because it gets used loosely across the category — often as a headline with the conditions left out.
It only applies when you’re running a local model. Point BreatheReply at Ollama or LM Studio — no key needed — and the entire thing happens offline: your message goes to localhost, the reply comes back, and zero tokens ever touch the network. Screenshot a group chat, write the reply, stay online or off — it makes no difference, because there’s no request to make.
It does not apply when you use a cloud model with your own key. In that case your text does leave your machine — it goes exactly where you pointed it. What doesn’t happen is us receiving it. We’d rather tell you that clearly than let you assume the strong version applies everywhere.
The honest summary is three sentences, and you can hold us to each one:
- Your words never reach us.
- Your words go only where you point them.
- Run a local model, and zero tokens leave your machine.
How to check a privacy claim yourself
You shouldn’t have to take our word for it — and you shouldn’t take any vendor’s. Four checks that work on any AI writing tool:
- Whose key is it? If you can’t supply your own, then the vendor is the one calling the model, and your text passes through their account regardless of what the policy page says.
- Is there a plain sentence about whether they can see your text? “We take privacy seriously” is not an answer. The honest ones say it outright: yes, or no.
- Does it still work with no cloud at all? This one is easy to verify — point it at a local model, disconnect from the network, and see whether anything happens. Tools without a real local path can’t pass this test, whatever their marketing says.
- Is the claim bounded? “Nothing ever leaves your device” is unfalsifiable for any tool that also supports cloud models. Claims that come with conditions attached are usually the ones being careful, not the ones being weak.
That last point is the reason this article exists. We’d rather be believed for the sentence we can defend than admired for one we can’t.
Answered plainly
What does the app itself send?
Nothing on its own. Two things can go out, and both are things you trigger:
- Usage analytics are off unless you turn them on in Settings → Privacy. If you opt in, the app records coarse signals — which mode you used, roughly how long a message was — and never the message itself. Leave it off and nothing about your usage is sent.
- Provider profiles are a small table of how each service expects to be called. Ours is fetched from our own site, and only when you press “Fetch latest profiles” in Settings — never in the background. It contains no message of yours and nothing about you.
Neither one carries your text. The only thing that ever sees your words is the model you pointed the app at.
On the website side, the only thing we collect is the email you leave in the waitlist form, used for one message at launch — the full list, retention and unsubscribe terms are in our privacy policy.
Will a local model be worse?
That depends on which model you run and on your machine — and it’s a trade-off we’d rather leave with you. Running locally is an option we support, not a claim that it’s the better answer for everyone. You can switch back to a cloud model in Settings at any time; we won’t decide the trade-off for you.
Privacy marketing tends to stay abstract, which is convenient — because replies aren’t abstract. The message you’d least want leaked is almost always the message you’d most want help with: the tense negotiation, the family friction, the thing you can’t say to your boss’s face. It’s the same weight another post tries to account for: the twelve minutes before you hit send.
A reply assistant you can’t trust with those is useless at exactly the moments it claims to help. So the trust came first, and the three sentences above are the whole arrangement.
Reply calmly. And privately.
Still deciding? The two approaches are laid out side by side here — including the cases where pasting into a cloud assistant is genuinely the better choice.
If that arrangement is the thing you were looking for, BreatheReply is the tool it came from: join the waitlist.